acceptodds
Under review as a conference paper at ICLR 2027

Reachability Is Not Enough: Selection Failure and Repair in Conditional Flow Matching

Abstract

Conditional flow matching (CFM) learns a velocity field by regression, but success is judged by the distribution it generates. If a model class contains a correct generator, must population CFM select one? It need not. We construct a finite-dimensional affine class for independent straight-line CFM with Gaussian endpoints in which every admissible positive-definite matrix weight measurable from the regression input excludes the correct generator. For identifiable affine classes, we characterize when reweighting can select a prescribed field under specified information access and derive the exact minimum global spectral contrast required. Near this obstruction, even the best attainable asymptotic variance for the weighted-regression ratio estimator studied here diverges, whereas an endpoint-derived moment equation using the same observations remains regular even at the obstruction. Controlled finite-update experiments show CFM moving correct synthetic and pretrained generators away from their targets. In a structured family, CFM attains lower regression risk than likelihood fitting yet much worse endpoint KL; in a richer version, endpoint fitting substantially reduces CFM's KL under a prespecified 10% empirical regression-risk budget.

Primary area:
generative models
Keywords:
conditional flow matching, generative modeling, model misspecification, surrogate objectives, loss reweighting, statistical efficiency, estimating equations, coupling design

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8fpeOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates whether the existence of an endpoint-correct generator within a model class is sufficient for conditional flow matching (CFM) to select such a generator. Its main contributions are:

  • Constructing an affine family with Gaussian distributions in which a correct generator exists, but every admissible positive-definite input-measurable weight excludes it.
  • Characterizing whether reweighting can select a prescribed velocity field under specified information restrictions.
  • Deriving the minimum global spectral contrast required for selection.
  • Analyzing the asymptotic variance of particular weighted-regression estimators near the selection obstruction.
  • Constructing an endpoint-derived moment equation that remains regular in the family under consideration.
Critical strengths

1. rigorous distinction between representability and objective-based selection

  • The main construction starts with a model class containing an endpoint-correct generator.
  • The result goes beyond showing reference-velocity misspecification.
  • It excludes every endpoint-correct generator in the class under the stated weight restrictions.
  • Allowing state-dependent positive-definite matrix weights makes the result stronger than a counterexample to uniform weighting.
  • The positive endpoint-error bound distinguishes the obstruction from a merely formal discrepancy.

2. quantitative characterization of reweighting feasibility and cost

  • The paper characterizes selection of a prescribed candidate in affine parameter classes.
  • The minimum spectral contrast explicitly depends on the information available to the weight.
  • The analysis shows that state-dependent and time-only weights can incur different costs.
  • It distinguishes changes to the population objective from importance sampling that preserves the objective.
  • The comparison with an endpoint-derived moment equation clarifies that the variance obstruction concerns a particular estimator protocol, rather than every possible use of the same observations.
Critical weaknesses

1. The neural diagnostics do not sufficiently establish practical significance

  • Teacher-defined target. The pretrained experiment replaces the real CIFAR-10 distribution with the teacher's Euler-100 distribution. This design appropriately guarantees a correct starting point, but does not directly represent learning from real image data.
  • Narrow model coverage. The image evidence relies on one checkpoint of one architecture. Two confirmation streams do not establish prevalence across teachers, architectures, or model scales.
  • Finite update budgets. Results after one and 200 updates demonstrate departure from the initial target, but do not establish discrepancy at a stable solution or convergence.
  • Observable drift does not quantify practical degradation. The Inception projections provide valid witnesses of distributional change, but do not quantify overall discrepancy or demonstrate deterioration in image quality.
  • The mechanism is not fully isolated. Sampled Adam updates, dropout, and numerical implementation all contribute to the experiment. The results do not establish population-gradient behavior in the neural model.
  • Statistical significance differs from practical significance. A narrow confidence interval around a small effect does not establish that the effect materially affects applications.
  • The scope is acknowledged, but significance remains limited. The abstract explicitly describes finite-update experiments, and the discussion acknowledges that convergence and prevalence remain open. Nevertheless, I find the neural evidence insufficient to support broader practical implications.

2. The connection between the theoretical constructions and broadly applicable repair remains limited

  • Construction-specific geometry. The strongest exclusion result relies on deliberately aligning the residual and fitting direction. It remains unclear how representative this geometry is of ordinary neural model classes.
  • A prescribed candidate is assumed. The reweighting characterization addresses selection of an already specified velocity field; it does not address how to identify a correct generator from data.
  • Incomplete temporal comparison. Table 2 uses a shared model class, but the temporal follow-up expands CFM while retaining likelihood fitting in the original family. The comparison between objectives in the expanded family remains incomplete.
  • Capacity has a substantial effect. Temporal expansion removes most of the base CFM KL. This makes it important to clarify the relative contributions of class restriction and objective selection to the empirical discrepancy.
  • Repair exploits known structure. The structured experiments use coordinates and features known to both learners. This is a valid controlled setting, but does not establish effectiveness when the structure must be learned.
  • Diagnostics rely on experimental ground truth. Teacher-derived KL and its decomposition are useful for post-hoc evaluation, but do not constitute a deployable diagnostic when the target distribution is unknown.
  • Scalability is not established. The computational savings from two endpoint corrections are evaluated on two 2D tasks. The 100D structured repairs also exploit special structure; corresponding evidence for neural-scale repair is absent.
  • Neural repair remains unconvincing. The exploratory companion improves some moment metrics but fails the structural-improvement criteria.
  • The risk budget is empirical. The result under a 10% regression-risk increase does not provide a population-risk-cap guarantee.
  • Implications for other methods require clarification. The paper does not establish how far the obstruction transfers to other paths, couplings, or objectives. This limits its applicability; it is not a requirement to prove failure across every generative framework.
Questions for authors
  1. Significance of the pretrained discrepancy: Beyond establishing distributional change, what do the current results show about the magnitude and practical consequences of the discrepancy? Is there existing evidence about its evolution under longer training or at a stable solution?
  2. Numerical and implementation controls: Which controls were performed specifically for the pretrained experiment? How large are the discrepancies attributable to precision, target generation/evaluation, and teacher-risk reconstruction relative to the reported effects?
  3. Novelty of the characterization: What is the essential new contribution relative to the metric-angle and constrained condition-number results discussed in Appendix A.4? Which part provides insight beyond the specific constructions?
  4. Capacity and objective: In the temporal follow-up, which conclusions can be attributed to objective selection when the methods are not fitted in the same expanded family? Which existing comparator helps separate capacity effects from objective effects?
  5. Deployability of repair: Which steps dominate computational cost as dimension and parameter count increase? Which assumptions about coordinates, features, or density tractability are essential?
  6. Scope of the weighting theory: How do the theorem's assumptions map onto practical time-weighting schemes? Is there a concrete setting in which the characterization yields a testable prediction?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 8fpe (1/4)

Thank you very much for the detailed review. We now report additional pretrained image metrics, longer-training observations and implementation controls. We also identify the existing comparisons that hold representation fixed and clarify the contribution beyond classical metric design. Convergence and broad neural applicability remain open.

Question 1. Magnitude and evolution of the pretrained discrepancy

The result is this. From a start that generates the target exactly, sampled CFM updates reach a model with lower fresh regression risk and a different generated law, in two independent streams. Regression risk therefore ranks an incorrect model ahead of a correct one. That is an objective mismatch at finite displacement. It does not settle convergence, prevalence or a material loss of image quality.

On magnitude, after one update the fresh risk falls by 3.3×10−53.3 \times 10^{-5} and 4.0×10−54.0 \times 10^{-5} per coordinate, and the fixed feature projection moves by 0.088 and 0.078. After 200 updates the risk falls by 5.8×10−45.8 \times 10^{-4} and 5.5×10−45.5 \times 10^{-4} (Table 3). The projection is a witness that the law changed and not a calibrated distance, and the directions for the two counts were fixed separately. Appendix F.2 states that the one-update changes are not evidence of a practically important loss of image quality.

On evolution, the submitted pretrained confirmations cover one and 200 updates; the separate exploratory run below extends to 2,000. Since submission, six 240,000-update runs in the synthetic neural setting retain lower held-out risk and detected endpoint discrepancies at all final raw and parameter-averaged checkpoints under fixed Heun-8 deployment. These are synthetic results, not longer CIFAR training, and establish neither stationarity nor convergence. We will provide the complete protocols and results in the consolidated follow-up.

Secondary image metrics use 10,000 generated images per model. FID uses CIFAR-10 training reference statistics. Teacher-reference FD is the same Frechet distance computed against teacher samples.

Each pair below lists confirmation 1, then confirmation 2.

  • Teacher: FID 6.6484, 6.3411; teacher-reference FD 4.9200, 4.8677.
  • Raw student, 1 update: FID 6.9083, 6.5528; teacher-reference FD 4.9628, 4.9003.
  • Raw student, 200 updates: FID 11.2689, 10.3479; teacher-reference FD 6.9417, 6.1999.
  • EMA, 200 updates: FID 6.6990, 6.3876; teacher-reference FD 4.9248, 4.8670.

The teacher-reference values for the teacher compare two independent teacher samples. They calibrate finite-sample variation and are neither a floor to subtract nor an equivalence threshold. All values are point estimates without confidence intervals. The raw 200-update models have worse recorded FID, and EMA strongly attenuates the change. The EMA starts at the teacher with decay 0.9999, and after 200 updates the teacher still has nominal weight 98.02% in the parameter average. This is not a mixture of generated distributions, and the short-horizon attenuation does not establish a general repair.

A separate exploratory run, outside the two confirmation streams, continued for 2,000 updates. Its real-data FID was 6.5337 for the teacher, 11.0391 after 200 updates, 10.6642 after 1,000 and 10.7398 after 2,000. The values at 2,000 updates follow.

Paired risk changes are per coordinate, in units of 10−510^{-5}. Brackets give nominal conditional 99% intervals for the fixed-witness change.

  • Raw student: risk change -69.75; witness change 0.7760 [0.7396, 0.8123]; real-data FID 10.7398; teacher-reference FD 6.2993.
  • EMA: risk change -22.82; witness change 0.1436 [0.1323, 0.1548]; real-data FID 7.0941; teacher-reference FD 4.9810.

The calibration between two teacher samples in this run is 4.9220. The EMA uses the same initialization and decay, and the teacher still has nominal weight 81.87% in the parameter average at 2,000 updates. The EMA witness is a secondary evaluation of the same fixed observable and not an additional primary confirmation. This run shows persistence over the measured horizon and supports no claim of stationarity or convergence.

AuthorsOpenReview ↗

Response to Reviewer 8fpe (2/4)

Question 2. Controls for the pretrained experiment

Appendix F.2 reports the following. Teacher and student are evaluated on identical fresh draws, with 16,384 triples for risk and 10,000 common sources for the endpoint. Target, training, risk and endpoint streams use separate seeds, and a pilot fixed the update counts and feature directions before the two confirmation streams. Dropout is active only in training.

The teacher's per-triple risks were reconstructed from the recorded evaluator and seeds. Combining each reconstructed difference with its baseline reproduces the retained student arrays byte for byte in all eight combinations of stream and checkpoint, and the appendix notes that this does not prove identity of every intermediate tensor. Separate CPU code reproduced the summaries, intervals and bootstrap results from the stored vectors, without rerunning GPU generation or feature extraction.

Pairing estimates the uncertainty of the difference between teacher and student, conditional on the fitted models. The nominal 99% intervals of Table 3 resolve that difference under the stated interval model. Their half-widths are 15% and 20% of the one-update risk change and 7% and 11% of the one-update projection change. They do not bound systematic precision error or variation across training runs.

Four further controls for the pretrained experiment give the following exploratory results, with per-coordinate risk changes in units of 10−510^{-5} as in Table 3. The witness intervals are nominal conditional 99% intervals for the fixed observable and are not additional primary confirmations.

  • Zero learning rate, 1 update: risk change 0.000; witness change 0.00000 [0.00000, 0.00000].
  • TF32 disabled, 1 update: risk change -1.906; witness change 0.07517 [0.06838, 0.08196].
  • Dropout disabled, 1 update: risk change -3.339; witness change -0.04612 [-0.05265, -0.03960].
  • Midpoint, 50 steps / 100 NFE, 200 updates: risk change -51.249; witness change 0.59968 [0.56772, 0.63164].

The zero-learning-rate control checks null behavior. Unchanged weights produce exactly zero paired risk and witness differences in this run. Departure also occurs with TF32 disabled. That run is not a float64 comparison and gives no systematic precision bound. Departure also occurs with dropout disabled, where the fixed observable moves in the opposite direction. That run also uses different training and evaluation seeds, so it does not isolate dropout as the cause of the sign reversal. The midpoint run defines its target with the midpoint teacher and uses the same deployment for the refit. It is a second matched construction of teacher and deployment and not an evaluation against the original Euler-100 target. The one-update and 200-update witness directions were fixed separately in the pilot.

Question 3. Novelty relative to Appendix A.4

The classical results cover two special cases. In the deterministic one-direction case, Theorem 2 reduces to the metric-angle threshold of the generalized Wielandt inequality. For finitely many information atoms, the problem is an instance of convex condition-number design and is solved numerically. Three things are new. The first is a closed form for measurable weights under an information restriction, (1+ρH)/(1−ρH)(1 + \rho_H)/(1 - \rho_H), with attainment below the boundary and exclusion at ρH=1\rho_H = 1 without any bound on the condition number. The second is the role of information. More information can only lower the required contrast, and in the squeeze family the difference is one of order, ε−2\varepsilon^{-2} against ∣ε∣−3\lvert \varepsilon \rvert^{-3}. The third is the use. The classical results do not ask whether regression selects a generator with the correct endpoint, which Theorem 1 answers for a whole correct set. Theorem 2 applies to any identifiable affine class satisfying its stated assumptions, with the reference target determined by the chosen coupling and path. The resulting lesson is that capacity to generate the target does not guarantee correct regression selection. The information available to the weight determines whether a prescribed correct field is selectable and at what contrast.

AuthorsOpenReview ↗

Response to Reviewer 8fpe (3/4)

Question 4. Capacity and objective

Your reading of Table 5 is correct. It cannot show that likelihood would beat CFM when both are optimized over the expanded family, and the paper says so. Three submitted comparisons do separate capacity from objective. Table 2 compares the two objectives in the same 6,725-parameter family, with 57.06 nats for CFM and 0.105 for likelihood. The original likelihood fit, at 0.104 nats, is a member of the expanded family because constant coefficients are allowed, while CFM selects a member at 6.49 nats with lower regression risk. Tables 6 and 7 fix the latent and contraction profiles of the temporal fits and vary only the drive, which carries 107,456 of the 107,600 parameters. There the regression-selected drive gives 6.515 nats, a drive fitted to endpoint samples gives 1.960 nats under the empirical 10% risk budget, and an oracle drive attains 0.227 nats, which shows substantial remaining approximation capacity in the same subfamily. Temporal capacity thus removes 88.6% of the CFM KL, and a change of criterion within the expanded family removes most of the remainder. This oracle uses teacher information, extreme coefficients and very high regression risk; it is a capacity diagnostic, not a practical repair. These are finite-data fits, and the membership statement is not a likelihood optimization over the expanded family. Our reply to Reviewer CEeF (item 4) discusses time weighting and finer temporal classes.

On theoretical reach, Appendix E already treats infinitely many correct generators and radial endpoints that need not be Gaussian. We have subsequently proved exclusion of every correct generator in a class allowing unrestricted spatial dependence and unrestricted time dependence on half the interval, with the halves linked by an explicit temporal constraint. This extension uses independent Gaussian endpoints with the same law and positive scalar input weights. The temporal constraint prevents representation of the reference velocity. We will present this extension and a complementary positive result, with their assumptions and proofs, in the consolidated follow-up.

Question 5. Deployability of the repairs

The corrections of Section 6.3 need target samples, a differentiable sampler and a feature class for the endpoint objective. Each correction computes the Jacobian of the feature residuals with respect to all parameters by forward sensitivities and then solves a dense least-squares problem. Both steps grow with the parameter count, and a neural-scale version would need reverse-mode differentiation through the sampler and matrix-free solves. We have not built or timed one, and Section 6.3 says that its timings do not establish neural-scale costs. The drive fits of Section 6.2 are linear least squares. They need the known coordinates, the known features and the triangular structure that makes the conditional mean linear in the drive. A tractable density is needed for the likelihood fits and for the KL evaluation, and the sample-only drive fits do not use it. The endpoint equation of Theorem 4 needs linear fields and identifies the endpoint only within the squeeze class.

Reviewer_ocHGOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the discrepancy between endpoint realizability and population-risk minimization in Conditional Flow Matching (CFM). The authors construct finite-dimensional affine velocity-field classes in which the target distribution is exactly reachable, yet the CFM objective fails to select a correct generator, even under arbitrary admissible positive-definite input-dependent reweighting. They further characterize the conditions under which reweighting can recover a correct generator and analyze the associated statistical costs. To address this limitation, the paper explores endpoint-based estimation and correction methods. Experiments on synthetic distributions, structured high-dimensional models, and a pretrained image generator demonstrate that lower CFM regression loss does not necessarily translate into better endpoint distribution matching, while endpoint-oriented corrections can substantially reduce the discrepancy.

Critical strengths

The paper provides a rigorous theoretical characterization of selection failure in CFM, demonstrating that even when a correct generator exists within the model class, positive-definite reweighting may fail to select it. The analysis further establishes precise conditions for successful reweighting and quantifies its statistical costs.

Critical weaknesses
  1. From the perspective of the continuity equation, it is not surprising that the orthogonal projection of the unrestricted (L^2)-optimal velocity field onto a restricted affine function class may fail to preserve the desired marginal distribution evolution, i.e., ∇⋅ρ(v−v⊥)≠0\nabla\cdot\rho(v-v^\perp)\neq0. Consequently, the projected velocity field need not generate the correct target distribution. Therefore, the fundamental observation of selection failure appears relatively straightforward rather than counterintuitive.

  2. The main theoretical results are established for highly restricted affine velocity-field classes, and it remains unclear whether similar selection failures occur in modern, expressive neural flow models. Moreover, the pretrained image-model experiments only demonstrate changes in the generated distribution after a limited number of CFM updates, rather than a significant degradation in generation quality at convergence. The proposed repair strategies are also primarily validated on structured synthetic models, leaving their scalability and effectiveness in practical generative modeling unclear.

Questions for authors
  1. The main impossibility result relies on a highly restricted affine velocity-field class. Does the whole-set selection failure persist when the function class is enriched, for example, by introducing additional independent temporal basis functions?
  2. If similar selection failures occur in neural flow models, can the proposed theoretical framework provide guidance on how to modify network architectures or parameterizations to avoid such failures? In particular, can the theoretical results identify structural properties of the model class that are sufficient to ensure the selection of an endpoint-correct generator?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ocHG (1/2)

Thank you for the review and for two questions that led to new results. The supporting note under the general comment gives the new statements with their arguments.

1. Selection failure and projection mismatch (Weakness 1)

We agree with the observation, and the paper starts from it. CFM selects the best L2(μ)L^2(\mu) approximation to the reference velocity uu, and such a projection generally does not preserve the reference marginals. Two parts of the paper go beyond this.

Leaving the reference path is different from missing the target. The condition ∇⋅[pt(v−u)]≠0\nabla \cdot [p_t (v - u)] \ne 0 says that vv does not follow the reference marginals. It does not say that the endpoint of vv is wrong. In the squeeze class the correct generator v∗v_\ast itself satisfies this condition. It takes a different route and reaches the target exactly (Figure 1). In the scale example of Section 2, no member of the class follows the reference marginals, and yet the time weight 1/Dσ1/D_\sigma makes regression select the correct member. The continuity-equation criterion is thus violated both where a simple reweighting repairs the selection and where no reweighting can. It cannot tell the two cases apart.

Theorems 1 and 2 answer the question it leaves open. Theorem 1 shows that no positive-definite weight measurable from time and state selects the correct generator, however anisotropic or ill-conditioned the weight is, and it gives a common endpoint-error bound. Theorem 2 gives, for affine classes, the exact contrast (1+ρH)/(1−ρH)(1 + \rho_H)/(1 - \rho_H) that a selecting weight needs as a function of the information it may use. Neither follows from failure of the unweighted projection, as the scale example shows.

The ingredients are classical, as Section 4 and Appendix A.4 acknowledge. The contribution is the question of whether containing a correct generator suffices for regression to select one, and its exact answer in these classes.

2. Scope of the theory and experiments (Weakness 2)

We agree with this description of the scope, and the general comment separates the three levels of evidence. Theorem 1 concerns one constructed class, and Theorem 2 makes no statement about how often the obstruction occurs. The finite-update experiments show that, from a start that generates the target exactly, sampled CFM updates reach a model with lower fresh regression risk and a different generated law. This is an objective mismatch at finite displacement. It does not settle convergence, prevalence or a material loss of image quality. Our replies to Reviewers CEeF (item 3) and 8fpe (Question 1) add image metrics for the pretrained experiment, against real CIFAR-10 and against the teacher, with EMA results and an exploratory run of 2,000 updates. The repairs are demonstrations in structured families.

Since submission, six 240,000-update runs in the paper's synthetic neural setting retain lower held-out CFM risk and detected endpoint discrepancies at all final raw and parameter-averaged checkpoints under fixed Heun-8 deployment. These extend the finite-horizon evidence without establishing convergence.

The theoretical scope also extends beyond the squeeze construction. Appendix E already treats infinitely many correct generators and radial endpoints that need not be Gaussian. We have subsequently proved exclusion of every correct generator with unrestricted spatial dependence and unrestricted time dependence on half the interval, subject to an explicit relation between the two halves. This extension uses independent Gaussian endpoints with the same law and positive scalar input weights; the temporal restriction prevents representation of the reference velocity. We will present it and a complementary positive result, together with the full experimental results, in the consolidated follow-up.

AuthorsOpenReview ↗

Response to Reviewer ocHG (2/2)

3. Additional temporal basis functions (Question 1)

Enlargement does two things. It can introduce new correct generators, and it adds ways to adjust every existing generator. A regression minimum must have zero loss derivative along every available adjustment. More correct candidates can help, and more fitting equations can exclude candidates that previously minimized the loss.

The squeeze family makes both effects explicit (supporting note, Items 1 and 2). Write e=v∗−ue=v_\ast-u and k=t+1/2k=t+1/2.

  • Transverse features are pointwise orthogonal to ee. Under a positive scalar weight they cannot cancel it and only add squared error. Regression leaves their coefficients at zero, so the original exclusion and error bound survive.
  • Adding ktekte to the original direction keke introduces correct generators vc=v∗+c(t2−t−3/4)ev_c=v_\ast+c(t^2-t-3/4)e. The new profile is symmetric about mid-time. In this construction, the second half of the flow therefore undoes the first, returning to the starting law, which is also the target. For 1<c<4/31 < c < 4/3, the residual multiplier changes sign across time, and a bounded positive time weight balances both normal equations. This is the escape.
  • Adding kt2ekt^2e leaves the correct set unchanged. But each correct member's residual is now itself an available fitting direction after multiplication by positive kk. Moving against it lowers loss for every admissible weight, excluding the whole line again.
  • Adding kt3ekt^3e introduces new selectable correct generators, as does every higher degree. The original generator stays excluded throughout (Appendix A.3).

Thus the first addition creates new correct candidates that can satisfy the fitting equations, and the second adds an equation that excludes them without adding new correct candidates. These are results for the stated feature families, not arbitrary enlargements. Exact selection and error size are also different questions. Unweighted population CFM has endpoint error W22W_2^2 of 0.1570.157, 2.3×10−42.3 \times 10^{-4}, 1.7×10−41.7 \times 10^{-4} and 5.0×10−75.0 \times 10^{-7} with zero to three aligned features, and the error tends to zero along the hierarchy.

4. Structural properties that ensure selection (Question 2)

The results give three sufficient properties.

  • The class contains the reference velocity. A global population minimizer under a fixed admissible positive input-measurable quadratic weight then reproduces it on the reference path and generates the target under the usual flow assumptions.
  • The minimizer admits all local gradient variations. Appendix H shows that such a minimizer preserves the reference marginals under positive scalar time weights and the regularity conditions stated there.
  • The fitted directions represent the endpoint conditions. This is new (supporting note, Item 3). In the squeeze family, one class with three parameters contains two directions whose normal equations are exactly the two endpoint conditions, and ordinary unweighted CFM is then endpoint-correct. Another class with three parameters excludes every correct generator under all admissible weights.

The number of parameters therefore does not decide the outcome. In this comparison, the decisive property is that the fitting equations enforce the endpoint conditions. In the squeeze family these directions can be written in advance, because the endpoint conditions are linear in the time profile. In a general nonlinear model they depend on the fitted flow, so we do not claim an architectural recipe. In the squeeze class, Theorem 4 provides another route by enforcing endpoint correctness through an estimating equation built from the same tuples.

Please tell us if either answer leaves part of your question open.

Thank you for the detailed responses and additional results. The new analysis of model-class enlargement substantially clarifies the scope of the selection failure, and the proposed sufficient structural conditions provide useful insights into when correct selection can be guaranteed. The additional experiments also strengthen the empirical evidence.

However, the theoretical extensions are still established for specific function classes, and the proposed structural conditions do not yet provide a general architectural guideline for expressive neural models. Moreover, the empirical evidence does not fully establish the prevalence or convergence behavior of selection failure in practical neural flow models.

Overall, the rebuttal addresses most of my concerns and clarifies the contributions and limitations of the paper. I will maintain my rating of 3 (Weak Accept).

AuthorsOpenReview ↗

Population selection in a finite neural class

Thank you for your follow-up comment. We have added results addressing your questions about population selection in neural classes and structural conditions for correct generation.

For a twofold stretch of a spherically symmetric source with full support and finite second moment, a state-only, bias-free ReLU network with a linear skip contains the exact generator at every width. Yet on the independent straight-line path, the global population minimizer of unweighted CFM has a Wasserstein error of 22.98% of the target's root-mean-square radius. The source need not be Gaussian. This tells us what exact minimization selects in that class; whether training reaches it is a separate question.

The note "Population CFM in finite neural classes" gives the proof and a complementary constructive result. For identical source and target laws and the specified temporal profiles, replacing fitted heads N1,N2N_1,N_2 by (N1−N2)/2(N_1-N_2)/2 and its negative makes the flow retrace its steps, enforcing the correct endpoint without increasing population risk. The correction stays within the same width. "State-only fields for Gaussian scaling" examines how spatial capacity and the interpolation path affect endpoint error when time is unavailable to the model.

Our general comment, Follow-up results: time weighting, longer neural training, and model capacity, reports the completed structured time-weighting comparisons and sustained synthetic-neural experiments, alongside the revised manuscript and self-contained supplement.

These results provide concrete design examples in specified classes. We agree that a general architectural guideline, training convergence, and the prevalence of selection failure in practical neural flow models remain open.

Reviewer_CEeFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies whether CFM selects an endpoint-correct generator whenever the model class contains one. The authors construct an affine Gaussian family in which every admissible positive-definite weight measurable from the regression input excludes the unique correct generator. They further characterize the minimum global spectral contrast required for reweighting to select a prescribed field, establish estimator-specific variance costs near the selection obstruction, and construct an endpoint-derived estimating equation that remains regular. Experiments demonstrate regression-risk reduction accompanied by endpoint drift, as well as improvements from endpoint-directed fitting in structured models.

Critical strengths

The paper identifies a meaningful distinction between endpoint realizability and reference-velocity realizability. A model class may contain a correct generator without containing the reference marginal velocity, and the best regression approximation need not be endpoint-correct. This distinction is clearly formulated and relevant to understanding restricted-model training objectives.

The whole-set exclusion result is stronger than simply exhibiting velocity misspecification or an unsuccessful weighting scheme. In the squeeze construction, all admissible positive input-measurable matrix weights select an incorrect parameter, with a common positive endpoint-error bound. The proof mechanism based on the positive alignment between the residual and the fitted direction is particularly interpretable.

The characterization of minimum global spectral contrast provides a quantitative extension beyond the impossibility result. The dependence on information access, and the distinction between state-dependent and time-only metrics, give the analysis additional substance.

The paper carefully separates population selection, sampling variance, finite-update behavior, and numerical deployment. It also explicitly acknowledges the limitations of its endpoint surrogates, including a negative neural experiment. This makes the scope of the evidence easier to assess.

Critical weaknesses

The main design relies on a structured, low-dimensional parameterization. In particular, the fitted direction is aligned with the residual at the correct generator, and the shared parameter prevents local adjustments across time and state. The paper establishes an important existence result, but provides limited evidence about when comparable obstructions arise in standard neural flow architectures. The discussion of admissible local variations in Appendix H is useful, but deserves greater prominence in explaining the boundary of the result.

The exact reweighting characterization concerns selection of an already specified candidate field in an affine parameter class. Although the relevant quantities are identifiable from CFM tuples for a fixed candidate, the result does not directly provide a procedure for finding an unknown endpoint-correct generator. The practical gap between certifying a candidate and discovering one could be explained more explicitly.

The pretrained CIFAR-10 experiment demonstrates departure from a teacher-defined target distribution, rather than deterioration relative to the actual CIFAR-10 distribution. Changes in selected Inception observables establish that the deployed law changes, but do not establish that perceptual quality worsens. Moreover, finite sampled Adam updates do not characterize the population minimizer or eventual convergence. The paper acknowledges these distinctions, but this experiment therefore provides a narrower bridge to practical image generation than the theoretical motivation might suggest.

The strongest endpoint-repair results exploit substantial known structure, including coordinates, features, and density factorization. The temporal expansion experiment also shows that increased modeling freedom removes most of the original CFM KL gap. A more systematic comparison across temporal capacity, time reweighting, and endpoint correction would help clarify when changing the objective is preferable to improving reference-velocity approximation. This would strengthen the practical interpretation without being necessary for the existence theorem itself.

The repair mechanisms do not yet constitute a general scalable solution. Determinant matching identifies the endpoint only within the constructed family, while the exploratory neural moment penalty improves covariance without satisfying the structural criteria. Consequently, the strongest contribution is the theoretical diagnosis; evidence for broadly useful repair remains preliminary.

Questions for authors

How robust is the whole-set exclusion result to enlarging the model class? For example, can a nontrivial endpoint-error floor be retained after adding several independent temporal or spatial features, rather than perturbing only the existing fitted direction?

Can the authors provide a practically estimable diagnostic for detecting selection failure when neither the correct generator nor its endpoint density is known? What additional assumptions would make the fixed-candidate characterization operational in this setting?

In the structured 100D experiment, how much of the remaining gap after temporal expansion can be removed by simple time-only weighting or additional temporal capacity? This comparison would help distinguish the benefits of endpoint fitting from those of better reference-field approximation.

For the pretrained image experiment, can the authors report distributional metrics against both the teacher-generated target and real CIFAR-10 data? These would distinguish departure from teacher fidelity from changes in actual data quality.

Could the main text more clearly distinguish three levels of evidence: whole-set population exclusion in the constructed family, local finite-update departure in neural models, and endpoint improvement under structured fitting?

4 comments
AuthorsOpenReview ↗

Response to Reviewer CEeF (1/3)

Thank you for your careful review and constructive questions. We answer in the order of your weaknesses and questions. The supporting note under the general comment gives the new statements with their arguments.

1. Enlargement and the boundary of Theorem 1 (Weakness 1, Question 1)

We agree that Theorem 1 is an existence result for a deliberately aligned class. Its boundary is as follows.

  • A fixed candidate does not become easier to select when features are added, since features only add normal equations (Appendix A.3). The original generator therefore stays excluded in every enlargement that keeps its fitted direction.
  • Additions transverse to the residual, meaning pointwise orthogonal to it, preserve the original exclusion and error bound under positive scalar weights. This holds for any number of features, with profiles that may depend on time and on the state (supporting note, Item 1). It does not hold for unrestricted matrix weights, where one transverse direction already removes the common bound.
  • Residual-aligned polynomial additions behave differently (Item 2). With one added feature, new correct generators appear and bounded positive time weights select some of them. With two, the correct set is unchanged and excluded again, and the smallest error that weights can approach is 2.5×10−62.5 \times 10^{-6}, against 0.0850.085 in the original class. With three or more, a selectable correct generator exists at every degree.

The reason is directional. Under a scalar weight, an orthogonal feature cannot cancel the existing residual, and using it only adds squared error. The first aligned temporal feature instead permits a correction whose second half undoes its first. The endpoint stays correct because source and target coincide. At some new correct generators, positive weights balance the regression derivatives across time. The next feature supplies a loss-decreasing adjustment at every member of that correct line without introducing a new correct generator. Selection is lost again.

Exact exclusion does not mean a large error. The original floor survives the transverse scalar-weight enlargement, but not the aligned hierarchy, along which unweighted CFM's error tends to zero. These enlargement results do not give a general classification for arbitrary state-dependent additions.

Appendix H gives the complementary positive statement. Under positive scalar time weights and the regularity conditions stated there, a population minimizer that admits all local gradient variations preserves the reference marginals. The squeeze class has no such freedom. The revision will state this boundary in the main text.

Appendix E already goes beyond the squeeze construction, with infinitely many correct generators and radial endpoints that need not be Gaussian. In subsequent work, we have also proved exclusion of every correct generator in a class allowing unrestricted spatial dependence and unrestricted time dependence on half the interval, with the two halves linked by an explicit temporal constraint. The result uses independent Gaussian endpoints with the same law and positive scalar input weights. The temporal constraint prevents the class from representing the reference velocity. We will present this extension and a complementary positive result, with their assumptions and proofs, in the consolidated follow-up.

AuthorsOpenReview ↗

Response to Reviewer CEeF (2/3)

2. Certifying a candidate and finding one (Weakness 2, Question 2)

Theorem 2 characterizes whether reweighting can select a supplied field; it neither finds an endpoint-correct field nor establishes its endpoint correctness. Theorem 1 instead excludes the whole correct set in its constructed class.

For a fixed candidate in an affine class, the selection quantities are identified from observable CFM tuples for both time-only and state-dependent weights (Section 4 and Appendix A.3). No separate estimate of the reference velocity is required. The practical difficulty is estimating the conditional moments, which with full state information is a regression on the full state, and obtaining reliable contrast bounds near the obstruction. For time-only weights the moments are functions of time alone. For nonlinear models the analogous calculation concerns tangent stationarity, not global selection.

On the additional assumptions you ask about, three separate things are needed. A candidate has to be supplied by another route, for example an endpoint-directed fit in the same affine class. Its endpoint correctness has to be established independently, because the characterization concerns selection and not the endpoint. The selection quantities then have to be estimated from fresh tuples, with the difficulties just described. Given all three, the characterization says whether any admissible reweighting of regression could select that candidate, and at what contrast. A candidate alone does not solve the practical problem.

When no candidate is known, what we can offer is held-out evidence of endpoint improvement within the same class. Starting from the CFM fit, an endpoint-directed correction in the same class is compared with it on independent samples. Section 6.3 does this with the unbiased estimator of energy distance. A supported improvement shows that the class contains a member with a better endpoint, in that discrepancy, than the fitted CFM model. It does not show that the fitted model is the population minimizer, or that the gap comes from population selection alone. We do not provide a general diagnostic for selection failure.

3. The pretrained experiment and image metrics (Weakness 3, Question 4)

Your description of the design is accurate. The teacher-defined target is what makes the start correct. The result is that, from a start that generates the target exactly, sampled CFM updates reach a model with lower fresh regression risk and a different generated law. Regression risk therefore ranks an incorrect model ahead of a correct one at finite displacement. The result does not settle convergence, prevalence or a material loss of image quality, and Appendix F.2 says that the one-update changes are small.

The two comparisons you ask for measure different things. Distance to the teacher's law measures departure from the correct start. Distance to real CIFAR-10 measures data quality, and it can move in either direction, because the teacher's own law differs from the data distribution.

Secondary image metrics use 10,000 generated images per model. FID uses CIFAR-10 training reference statistics. Teacher-reference FD is the same Frechet distance computed against teacher samples.

Each pair below lists confirmation 1, then confirmation 2.

  • Teacher: FID 6.6484, 6.3411; teacher-reference FD 4.9200, 4.8677.
  • Raw student, 1 update: FID 6.9083, 6.5528; teacher-reference FD 4.9628, 4.9003.
  • Raw student, 200 updates: FID 11.2689, 10.3479; teacher-reference FD 6.9417, 6.1999.
  • EMA, 200 updates: FID 6.6990, 6.3876; teacher-reference FD 4.9248, 4.8670.

The teacher-reference values for the teacher compare two independent teacher samples. They calibrate finite-sample variation and are neither a floor to subtract nor an equivalence threshold. All values are point estimates without confidence intervals. The raw 200-update models have worse recorded FID, and EMA strongly attenuates the change. The EMA starts at the teacher with decay 0.9999, and after 200 updates the teacher still has nominal weight 98.02% in the parameter average. This is not a mixture of generated distributions, and the short-horizon attenuation does not establish a general repair.

Separately, in Section 6.1's synthetic neural setting, six 240,000-update runs retain lower held-out CFM risk and detected endpoint discrepancies at all final raw and parameter-averaged checkpoints under fixed Heun-8 deployment. These extend the finite-horizon evidence; they do not establish convergence or describe sustained training of the pretrained CIFAR model. We will present the full protocols and results in the consolidated follow-up.

AuthorsOpenReview ↗

Response to Reviewer CEeF (3/3)

4. Capacity, time weighting and endpoint fitting (Weakness 4, Question 3)

The submission contains three comparisons that separate capacity from objective in the structured family. All are finite-data fits.

  • Same family, different objective. In the shared 6,725-parameter family, CFM reaches 57.06 nats and two-stage likelihood 0.105 nats, with CFM at lower regression risk (Table 2).
  • Membership. Constant coefficients embed the original family in the 16-interval family. The original likelihood fit, at 0.104 nats, is therefore a member of the expanded family, where CFM selects a member at 6.49 nats with lower regression risk (Table 5). Likelihood was not optimized over the expanded family.
  • Same expanded subfamily, different criterion. Tables 6 and 7 fix the latent and contraction profiles of the temporal fits and vary only the drive, which carries 107,456 of the 107,600 parameters. The regression-selected drive gives 6.515 nats, and a drive fitted to endpoint samples gives 1.960 nats under the empirical 10% risk budget. An oracle drive attains 0.227 nats in total, with conditional-mean KL about 0.020 nats, which shows substantial remaining approximation capacity in the same subfamily. This oracle uses teacher information and extreme coefficients and regression risk; it is a capacity diagnostic, not a practical repair.

Temporal capacity removes 88.6% of the CFM KL, and 6.28 of the remaining 6.49 nats are in the conditional-mean term.

The weighted comparison missing from the submission has now been completed on the same five archived datasets, with the same teacher and Euler-16 deployment. On common evaluation probes, sample inverse-spread weighting reduces mean base-family KL from 57.019 to 34.854 nats, but changes temporal-family KL only from 6.487 to 6.466. Oracle inverse-spread and logit-normal weights give 31.047 and 45.216 nats in the base family, and 6.545 and 7.026 in the temporal family. Thus these tested weights improve the base family but provide little or no improvement after temporal expansion. These are finite-data comparisons of particular weights, not exclusion under every admissible weight. The consolidated follow-up will provide the complete protocols and records.

Two population facts limit what the alternatives can do in the 16-interval family. For unpenalized population regression with independently parameterized and independently constrained intervals, multiplying each interval's loss by a positive constant leaves its minimizer unchanged. A time weight that is constant on each interval therefore does not change the fit. This does not automatically extend to a finite implementation with fixed ridge penalties or jitter. With the Euler-16 sampler fixed, an independently parameterized piecewise-constant refinement with the same coefficient ranges does not enlarge the set of deployed maps, because the sampler reads the field only at the sixteen grid times. In the Gaussian control, the limit of such refinement gives thickness 0.0691, against 0.0714 for 16 intervals and the target 0.0938. Item 4 of the supporting note gives an exact population comparison of capacity, reweighting and an endpoint constraint in the squeeze hierarchy.

5. Scope of the repairs and levels of evidence (Weakness 5, Question 5)

We agree. The repairs are demonstrations in structured settings, and the neural moment penalty is reported as a negative result. The general comment separates the three levels of evidence, and the revision will state it in the main text.

Please tell us which of these answers leaves a question open, and which control you would find most informative.

AuthorsOpenReview ↗

Follow-up on model capacity and the revised evidence

Thank you again for your review. Your questions motivated several additions to the revision. We would like to draw your attention to the completed experimental reports and two new supporting notes, summarized in our general comment, Follow-up results: time weighting, longer neural training, and model capacity.

On model enlargement (Question 1), Appendix I now contains the proofs discussed in our earlier reply. Complementing those results, a separate construction establishes an exact population result for a finite neural class. For a twofold stretch of a spherically symmetric source with full support and finite nonzero second moment, a state-only, bias-free ReLU network with a linear skip contains the exact generator at every width. Yet its unweighted population CFM optimum on the independent straight-line path has a Wasserstein error of 22.98% of the target's root-mean-square radius under exact ODE generation. The source need not be Gaussian. This establishes a consequential error at the global population optimum of a specified neural class, independently of whether training reaches it.

"Population CFM in finite neural classes" gives the proof. "State-only fields for Gaussian scaling" complements it by showing that enlarging spatial capacity can increase endpoint error, while changing the interpolation path can substantially reduce it. These results make the role of unavailable time information explicit; they do not establish prevalence in standard time-conditioned networks.

For the capacity-versus-objective comparison (Question 3), the revised Appendix F.3 and supplement now provide the full protocols and records behind the weighting results previewed in our earlier reply. Sample inverse-spread weighting changes temporal-family KL from 6.487 to 6.466 nats, whereas endpoint drive fitting within the fixed-profile subfamily reaches 1.960 under the empirical 10% risk budget. This compares particular procedures within their stated constraints, rather than establishing exclusion under every weight.

The revision also incorporates the requested teacher-reference and real-data image metrics, documents the six sustained synthetic-neural runs, and adds nine exploratory runs examining learning-rate decay and inactive units. These 15 training runs are synthetic and finite-horizon. They do not extend CIFAR training or establish convergence. In response to Question 5 and your Appendix H suggestion, the main text now distinguishes the three evidence levels and states the local-variation conditions for preserving the reference marginals.

The new theoretical notes are separate from the revised PDF. The general update identifies where each result, proof and reproducibility record can be found.

Comments on the paper

6 comments
AuthorsOpenReview ↗

General response: what the model can generate versus what CFM selects

We thank the reviewers for their careful and constructive reviews. The replies address individual questions. The supporting note under this comment gives the new mathematical results and arguments.

The contribution. We identify a limitation of CFM's population objective: even exact global minimization can exclude every correct generator in an endpoint-realizable model class. The key distinction is between capacity to generate the target and capacity to reproduce the chosen reference velocity. Our construction removes sampling, optimization and ODE discretization error. Every admissible positive-definite input-measurable quadratic loss weight still selects an incorrect generator, with a common positive endpoint-error bound (Theorem 1). Thus endpoint realizability alone does not guarantee correct CFM selection. For affine classes with nonredundant fitting directions, Theorem 2 characterizes when positive reweighting can select a specified field and the minimum contrast required, given the weight's information access. Near the squeeze obstruction, reweighting can restore correct population selection. Yet the specified regression estimator's asymptotic variance grows without bound as the obstruction is approached, even under the best positive selecting weights (Theorem 3). An endpoint-derived estimator using the same observations retains finite asymptotic variance at the obstruction (Theorem 4).

Why a correct generator can lose. The reference flow shrinks the Gaussian equally along both axes and then expands it back. The correct squeeze generator compresses one axis while stretching the other, turning circular contours into ellipses, then reverses the deformation. Both reach the target, but regression penalizes the squeeze generator's velocity mismatch. Within our class, a small adjustment toward the reference velocity lowers regression loss while spoiling the endpoint. This move is along −ke-ke, where e=v∗−ue=v_\ast-u and k>0k>0. Every admissible positive-definite weight WW rewards it, since ke⊤We>0k e^\top W e>0 wherever e≠0e\ne0. Reweighting cannot make the correct generator a minimum.

Three levels of evidence.

  • Population theory (Theorems 1 to 4). The results establish selection failure in constructed affine classes and characterize feasibility and statistical costs of repair. They raise questions for neural models: when does failure occur, which structural properties prevent it, and which interventions repair it?
  • Finite-update neural evidence (Section 6.1). Starting from a generator that produces the target exactly, sampled CFM updates yield lower fresh regression risk but a different generated law. This is evidence of objective mismatch at finite updates. Its implications for convergence and perceptual image quality remain open.
  • Structured endpoint fitting (Sections 6.2 and 6.3). Endpoint-directed fits reduce KL in the structured 100D family and energy distance in the 2D correction experiments. We do not claim a general scalable repair.

New in this response. The supporting note addresses class enlargement and structural conditions for correct selection through controlled extensions of the squeeze family. Enlargement can introduce new correct generators, but it also gives regression more ways to lower the loss at existing ones. A minimum must resist every such adjustment. More parameters therefore need not make exact selection easier. Two classes with three parameters have opposite outcomes: one excludes every correct generator under all admissible weights, while the other's unweighted normal equations enforce endpoint correctness. A targeted revision will include these results with proofs. The CEeF and 8fpe replies also present additional image metrics, EMA results and controls from the existing pretrained-experiment records.

Beyond the squeeze class. Appendix E's elliptic family has infinitely many correct generators and radial endpoints that need not be Gaussian. With independent endpoints, every admissible time-only matrix weight and every admissible positive state-dependent scalar weight excludes all of them. Separately, uniform-time CFM fails for every coupling of the fixed source and target. Both theoretical constructions have one fitted parameter. In the structured 100D family, the velocity is linear in 6,725 coefficients and the exact teacher belongs to the deployed class. The numerical CFM fit has lower fresh regression risk than the same-family likelihood fit, but forward KL is about 57 versus 0.1 nats (Table 2). This is a finite-data comparison with constrained coefficients and fixed Euler-16 deployment, not a population exclusion.

For practitioners, the immediate implication is to assess endpoint performance independently of CFM risk. Equation (17) formalizes a local mechanism for disagreement. A reliable diagnostic for general neural applications remains open.

We welcome follow-up questions on any points that remain unresolved.

AuthorsOpenReview ↗

Supporting note (1/2): how model enlargement changes CFM selection

All statements concern the squeeze family of Section 3 with tilt parameter ε=0\varepsilon=0, so its original fitting direction is keke, with exact ODE generation, source and target N(0,I2)N(0, I_2), independent straight-line CFM and the admissible weights of Theorem 1. They are new in this response. The arguments are short, and the revision will carry complete proofs.

Setup. Recall k=t+1/2k = t + 1/2, D=(1−t)2+t2D = (1-t)^2 + t^2, a=(2t−1)/Da = (2t-1)/D, v∗=C0xv_\ast = C_0 x, u=axu = a x and the residual e=v∗−u=Exe = v_\ast - u = E x. Consider fields v=v∗+g(t)ev = v_\ast + g(t) e for time profiles gg such that g(t)(2t−1)g(t)(2t-1) is bounded. Their coefficient matrices are diagonal, so the endpoint is a centered Gaussian whose two log standard deviations are −L1−L2-L_1 - L_2 and L1−L2L_1 - L_2, where

L1(g)=∫01g(t)(2t−1)dt,L2(g)=∫01g(t)a(t)dt.L_1(g) = \int_0^1 g(t) (2t-1) dt, \qquad L_2(g) = \int_0^1 g(t) a(t) dt .

The field is endpoint-correct exactly when L1=L2=0L_1 = L_2 = 0. Its regression residual is (1+g)e(1+g) e. For an admissible weight WW, the model-dependent part of the weighted risk is ∫01(1+g)2mWdt\int_0^1 (1+g)^2 m_W dt with mW(t)=E[e⊤We∣T=t]m_W(t) = \mathbb E[e^\top W e \mid T = t], which is positive for t≠1/2t \ne 1/2. In these classes every admissible matrix weight therefore acts as a positive time density. We write q(t)=E[∥e∥2∣T=t]=2(2t−1)2(D+1/D)q(t) = \mathbb E[\lVert e \rVert^2 \mid T = t] = 2 (2t-1)^2 (D + 1/D).

Item 1. Transverse features under positive scalar weights. Enlarge the class to

v=v∗+αke+k∑jβjϕjJe,v = v_\ast + \alpha k e + k \sum_j \beta_j \phi_j J e ,

where JJ is the rotation by a quarter turn and the ϕj(t,x)\phi_j(t, x) are finitely many scalar profiles of time and state, with finite weighted moments, a nonsingular regression Gram matrix and well-posed flows. Since e⊤Je=0e^\top J e = 0 and ∥Je∥=∥e∥\lVert J e \rVert = \lVert e \rVert at every point, every positive scalar weight w(t,x)w(t, x) gives the model-dependent risk

E[w∥e∥2((1+αk)2+k2(∑jβjϕj)2)].\mathbb E \Big[ w \lVert e \rVert^2 \Big( (1 + \alpha k)^2 + k^2 \big( \sum_j \beta_j \phi_j \big)^2 \Big) \Big] .

The minimizer has every βj=0\beta_j = 0 and the same α\alpha as in Theorem 1. The selected field, the interval −2<α<−2/3-2 < \alpha < -2/3 and the error bound of Theorem 1 are unchanged.

This statement needs scalar weights. With full matrix weights and one transverse direction kJek J e, the original generator stays excluded, but the common error bound does not survive. For each β≠0\beta \ne 0 the candidate field v∗+βkJev_\ast + \beta k J e has ρH=G2+β2H2/Qβ\rho_H = \sqrt{G^2 + \beta^2 H^2} / Q_\beta in Theorem 2, with G=∫kqdtG = \int k q dt, H=∫k2qdtH = \int k^2 q dt and Qβ=∫kq1+β2k2dtQ_\beta = \int k q \sqrt{1 + \beta^2 k^2} dt. This is below one, so a bounded positive matrix weight selects the candidate. Its endpoint law tends to the target as β→0\beta \to 0, while the required contrast diverges.

Item 2. Residual-aligned polynomial features. Let Cn\mathcal C_n be the class of fields v∗+kpev_\ast + k p e with pp a polynomial in tt of degree at most nn. Its fitted directions are ke,kte,…,ktnek e, k t e, \dots, k t^n e, so C0\mathcal C_0 is the original class and nn counts the added features. Write h=1+kph = 1 + k p for the residual multiplier of a member. It is a polynomial of degree at most n+1n+1 with h=1h = 1 at t=−1/2t = -1/2.

(a) Selection rule. A member of Cn\mathcal C_n is selected by some admissible weight exactly when hh changes sign n+1n+1 times inside (0,1)(0, 1). Stationarity under WW reads ∫mWkhtjdt=0\int m_W k h t^j dt = 0 for j=0,…,nj = 0, \dots, n. If hh changes sign at only s≤ns \le n interior points, the polynomial p0p_0 of degree ss that vanishes there makes hp0h p_0 of one sign, so ∫mWkhp0dt≠0\int m_W k h p_0 dt \ne 0 for every positive mWm_W. If hh changes sign n+1n+1 times, no nonzero polynomial of degree at most nn makes hph p of one sign. Then ρH<1\rho_H < 1 in Theorem 2, and a bounded positive time weight selects the member.

(b) Earlier fields. Every field of Cn\mathcal C_n is excluded in Cn+1\mathcal C_{n+1}. Its residual is heh e, and the next class contains the direction khek h e, because deg⁡h≤n+1\deg h\le n+1. Their weighted pairing is E[kh2e⊤We]>0\mathbb E[k h^2 e^\top W e]>0, so moving in the opposite direction lowers loss. Within this polynomial hierarchy, a field can be selected only in the smallest class that contains it. Escape therefore requires a newly introduced correct generator.

(c) Correct members. The functions 2t−12t - 1 and aa are odd about t=1/2t = 1/2, so both endpoint conditions involve only the part of hh that is odd about 1/21/2. For n=0n = 0 the only correct member is v∗v_\ast. For n=1n = 1 the single odd coefficient must vanish, which leaves the line gc=c(t2−t−3/4)g_c = c (t^2 - t - 3/4) with hc=1−c+c(t−1/2)2h_c = 1 - c + c (t - 1/2)^2. For n=2n = 2, write the odd part as u1r+u3r3u_1 r + u_3 r^3 with r=2t−1r = 2t - 1. The two conditions are u1/3+u3/5=0u_1/3 + u_3/5 = 0 and (2−π/2)u1+(π/2−4/3)u3=0(2 - \pi/2) u_1 + (\pi/2 - 4/3) u_3 = 0, with determinant (12π−38)/45≠0(12\pi - 38)/45 \ne 0, so the correct set is the same line. For n≥3n \ge 3 it is strictly larger at every step.

Continued in Supporting note (2/2).

AuthorsOpenReview ↗

Supporting note (2/2): how model enlargement changes CFM selection

(d) Classification. A selectable correct generator exists exactly when nn is neither 0 nor 2.

  • No added feature. Only v∗v_\ast is correct, and no admissible weight selects it (Theorem 1).
  • One added feature. The correct set is the line gcg_c, including v∗v_\ast at c=0c=0. Selection is possible exactly for 1<c<4/31 < c < 4/3.
  • Two added features. The correct set is the same line, and no admissible weight selects any member.
  • Three or more added features. New correct members appear at every nn, and selection is possible at every degree.

For n=1n = 1, the roots of hch_c are t=1/2±1−1/ct = 1/2 \pm \sqrt{1 - 1/c}, which lie in (0,1)(0, 1) exactly when 1<c<4/31 < c < 4/3. A selecting weight is then w=ω/kw = \omega / k, with ω=1\omega = 1 where hc≥0h_c \ge 0 and ω=λ\omega = \lambda elsewhere, where λ\lambda is the ratio of ∫qmax⁡(hc,0)dt\int q \max(h_c, 0) dt to ∫qmax⁡(−hc,0)dt\int q \max(-h_c, 0) dt. The first normal equation holds by the choice of λ\lambda and the second by symmetry about 1/21/2. For n=2n = 2, every correct multiplier has degree at most two, while three sign changes are required. For odd n=2m−1n = 2m - 1, take h=∏j(r2−aj2)/(4−aj2)h = \prod_j (r^2 - a_j^2)/(4 - a_j^2) with distinct a1,…,ama_1, \dots, a_m in (0,1)(0, 1). It is even, equals one at r=−2r = -2 and has n+1n+1 simple roots in (−1,1)(-1, 1). For even n=2m≥4n = 2m \ge 4, take h=cmrPm(r2)h = c_m r P_m(r^2), where PmP_m is the monic orthogonal polynomial of degree mm for dμ=s(1+s)−1dsd\mu = \sqrt{s} (1+s)^{-1} ds on (0,1)(0, 1) and cm=−1/(2Pm(4))c_m = -1/(2 P_m(4)). With s=r2s = r^2 the endpoint conditions become ∫(1+s)Pmdμ=0\int (1+s) P_m d\mu = 0 and ∫Pmdμ=0\int P_m d\mu = 0, which hold by orthogonality, and hh has n+1n+1 simple roots in (−1,1)(-1, 1).

(e) Exact selection and error size are different questions. Unweighted CFM's endpoint error W22W_2^2 is 0.1570.157, 2.3×10−42.3 \times 10^{-4}, 1.7×10−41.7 \times 10^{-4} and 5.0×10−75.0 \times 10^{-7} for n=0,1,2,3n = 0, 1, 2, 3. At n=2n = 2, where no weight selects a correct generator, the infimum of the error over all admissible weights is 2.53×10−62.53 \times 10^{-6}. It is positive and not attained, and the corresponding infimum in the original class is 0.0850.085. The proof reduces to one convex variable and will be in the revision. Unweighted CFM's error also tends to zero along the hierarchy. With s1,s2s_1, s_2 as in Item 3, the unweighted fit hnh_n has Lj=∫qhnsjdtL_j = \int q h_n s_j dt. Comparison with p=−∑i≤n(−r/2)ip = -\sum_{i \le n} (-r/2)^i, for which 1+kp=(−r/2)n+11 + k p = (-r/2)^{n+1}, gives ∥hn∥q≤Q02−(n+1)\lVert h_n \rVert_q \le \sqrt{Q_0} 2^{-(n+1)} in L2(qdt)L^2(q dt) with Q0=∫qdtQ_0 = \int q dt. Hence

W22≤2[exp⁡(Γ2−(n+1))−1]2,Γ=Q0(∥s1∥q+∥s2∥q)≈1.44.W_2^2 \le 2 \Big[ \exp\big( \Gamma 2^{-(n+1)} \big) - 1 \Big]^2, \qquad \Gamma = \sqrt{Q_0} \big( \lVert s_1 \rVert_q + \lVert s_2 \rVert_q \big) \approx 1.44 .

Item 3. Two classes with three parameters. Define the smooth fields Φ1(t,x)=−(K0+I/D)x/(2(D+1/D))\Phi_1(t, x) = -(K_0 + I/D) x / (2 (D + 1/D)) and Φ2=Φ1/D\Phi_2 = \Phi_1 / D, and consider the class

v=v∗+θke+αΦ1+βΦ2.v = v_\ast + \theta k e + \alpha \Phi_1 + \beta \Phi_2 .

In the notation of the setup, Φj=sje\Phi_j = s_j e with s1=(2t−1)/qs_1 = (2t-1)/q and s2=a/qs_2 = a/q. These scalar profiles are singular at t=1/2t = 1/2, but the fields are smooth and the profiles are square-integrable for qdtq dt. By construction Lj(g)=∫qgsjdtL_j(g) = \int q g s_j dt. For unweighted CFM, the normal equations in the two added directions read ∫qhsjdt=0\int q h s_j dt = 0, that is ∫h(2t−1)dt=0\int h (2t-1) dt = 0 and ∫hadt=0\int h a dt = 0. Since 2t−12t - 1 and aa integrate to zero, these are exactly L1(g)=L2(g)=0L_1(g) = L_2(g) = 0. The unweighted CFM minimizer is therefore endpoint-correct. It is a different correct generator from v∗v_\ast, which stays excluded. Its model-dependent risk is 0.00170.0017, against 0.1780.178 for the unweighted fit in the original class.

The class C2\mathcal C_2 also has three parameters, and no admissible weight selects a correct member there. The difference lies in whether the fitted directions represent the two endpoint conditions. In this family those conditions are linear in the profile, which is why the two directions can be written in advance.

Item 4. Capacity, reweighting and an endpoint constraint. For n=0,…,3n = 0, \dots, 3, we compare unweighted CFM, the best admissible reweighting and an endpoint-constrained fit. The latter minimizes unweighted CFM risk subject to L1=L2=0L_1=L_2=0, giving zero endpoint error in every case. Errors below are W22W_2^2.

Risk ratios below compare constrained with unweighted model-dependent risk.

  • n=0n=0. Unweighted error 0.1570.157; best reweighting has infimum 0.0850.085. Added risk: 36.6% of total risk; risk ratio 7.8.
  • n=1n=1. Unweighted error 2.3×10−42.3 \times 10^{-4}; best reweighting gives zero. Added risk: 0.011% of total risk; risk ratio 1.04.
  • n=2n=2. Unweighted error 1.7×10−41.7 \times 10^{-4}; best reweighting has infimum 2.5×10−62.5 \times 10^{-6}. Added risk: 0.24% of total risk; risk ratio 10.5.
  • n=3n=3. Unweighted error 5.0×10−75.0 \times 10^{-7}; best reweighting gives zero. Added risk: 0.0001% of total risk; risk ratio 1.05.

This is an exact population control in a constructed family. It is not evidence about the experiments of Section 6.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.